What you need to gather before hiring an AI chatbot (the 6-item folder)

Five hours of your own work. That is what it takes to gather the six items worth having on the table before the first meeting with a chatbot vendor, and none of the six is technical: the three numbers from a one-week count, the twelve questions you get asked most with a single official answer, the current price list with its exceptions, the escalation rule with a name and a schedule, the table of where each piece of data lives and who authorizes access, and who is currently listed as the owner of your company's Meta account.

Four of the six depend only on you and fit into one afternoon. The other two are not measured in your hours but in how many people have to agree: the price that was never written down because three partners decide it case by case, and the password held by the agency that built your website. Those two start today, along with the fifteen minutes it takes to find out whose name the Meta account is under, which is the piece of information that shows up at the last minute and under a former employee's name.

The requirement lists that circulate enumerate what the vendor solves: the number, the verification, the platform, the integration. None of that is gathered by you. What you do gather is the bot's raw material, and it is exactly what does not appear on any list: a price nobody wrote down, a password held by the agency, and a Meta account under a former employee's name.

The folder of 6, with the real work each one takes

• One-week count: three numbers. 25 minutes. You do it yourself.
• Twelve frequently asked questions with a single official answer. 1 hour, plus between two and four decisions that cannot be delegated.
• Current price list with exceptions and who authorizes them. One afternoon if there is one person who decides on the spot; several meetings if there are three partners who have to agree.
• Escalation rule written in three lines. 30 minutes, and only the person who decides can sign it.
• Table of where each piece of data lives and who authorizes access. 40 minutes to put together. What stretches is not building it: it is how many rows end up with someone outside.
• Who is listed as the owner of the Meta account and the number. 15 minutes to find out.

Almost five hours of your own time adding up the minimums. Four of the six you finish in one afternoon. The other two are not measured in your hours, they are measured in how many people have to agree.

The three numbers from the count keep them from quoting you an oversized plan

The method is in how many inquiries justify a chatbot: conversations from a normal week, extrapolated to the month. Do it once and write down three numbers: how many conversations came in, how many arrived outside your business hours, and how many of those never got an answer.

That post uses the three numbers to decide whether it is worth doing. Here they serve a different purpose, one that plays out inside the meeting: setting the size of what you will be quoted.

Conversation platforms are sold in volume tiers, and Meta bills per 24-hour conversation window, not per message. Published plans run from a few tens to a couple hundred dollars a month, and the step between one tier and the next is defined by the volume you declare. A business with 180 real monthly conversations that walks into the meeting saying "we get around 1,200 messages a month" is asking, without realizing it, for the quote of the tier above, and it will pay for that every month of the year.

The other two numbers are not decorative either. The out-of-hours one is what sizes the escalation coverage: if 70% of your inquiries come in after 7 p.m., the conversation about who handles those conversations changes tone. And the one for inquiries that never got an answer is the only one that defines a minimum verifiable scope: that, and not "improving customer service", is what the bot has to move for the project to be worth anything.

Twelve official answers, not sixty questions

The folder does not ask for a manual. It asks for twelve questions: the ones that concentrate the bulk of what comes in. Price, hours, whether there is stock or an appointment slot, payment methods, where you are located, whether you ship. You already know those by heart.

The work is not in the questions column: it is in the one next to it. Ask the two or three people who handle WhatsApp separately what they answer to each of the twelve, and write down their exact wording in separate columns. Out of twelve, between two and four will come back with two versions that do not match: one says shipping is free above a certain amount and the other says it is always charged; one asks for a 30% deposit and the other says "it depends on the job"; one accepts that health plan on Saturdays and the other does not.

Choosing which of the two versions becomes the company's is the real deliverable of this hour, and the vendor cannot do it. The bot will say one single version, to everyone, always. The moment you write it down, it stopped being the criterion of whoever is answering and became policy.

A price that is not written down does not exist for the bot

Of the six, this is the one that does not depend on your calendar but on other people's. Three columns: price, effective date, who can change it or authorize a discount.

What has to be written down is not the list, which probably already exists: it is the exceptions that today live in someone's head. The cash-payment discount. The different price for a health plan or for the customer who buys every month. The delivery surcharge by zone that gets estimated by eye. And the most common trap: the price published on the website does not match the one quoted over WhatsApp.

The proof that this part is finished is concrete and you can run it in two minutes: hand the sheet to someone who started working last week and ask them to answer "how much does it cost?" for three different cases, including one with an exception. If they have to ask someone, the sheet is not ready. If they answer all three, the vendor has what it needs to quote firmly and not in a range.

The escalation rule is written in three lines and signed by whoever decides

Three lines, not one more:

1. What the bot never answers. Do not copy a generic list: pull it from your own history. Look at last year's complaints that ended in a discount, a refund, or a call from the owner. Those topics are your list, and they are different from the ones at the business next door.
2. What triggers the escalation. A trigger has to be detectable in a text. The classic mistake is writing "escalate when the customer gets angry": that is not a rule, it is an intention, and there is no way to program it.
3. Who receives it, by first and last name, during which hours, and what happens outside those hours.

The third line is the only one of the six items in the folder that assigns work to a specific person, and that is why it is the only one you cannot delegate to an employee: whoever signs it has to be able to decide that this person drops what they are doing to handle a chat.

The check, before the meeting: send line 3 in writing to the person you named. If they answer "me?", the rule does not exist yet. On the other side of the counter, what you have to demand from the vendor about this is a different matter, and it is in the questions to ask before hiring them.

The data table and the question that predicts whether there is an integration

One row for each piece of data the bot will need to state or record: appointments, stock, prices, order status, customer details.

Column three is the one that predicts almost everything: do you get into that system from the browser with a username and password, or is it a program installed on a computer at the store? Through the browser, in the vast majority of cases there is a way to connect it. Installed program with no web version: no integration is possible and the scope is trimmed down to informing and taking the request, with someone entering it by hand afterwards. Shared spreadsheet: it connects without a problem, but it opens the other question, which is who keeps it up to date.

Column four is not counted in days: it is counted in signatures. Add up how many rows end with someone who is not in your company —the agency that manages the store, the previous vendor who never returned the token—. Zero third-party rows means the date depends on you. Two or more means the date depends on people who will not be in the meeting, and that is worth saying before somebody promises one.

Who owns your Meta account: fifteen minutes, today

Go into Meta Business Suite, Business settings, and look at one single thing: whether the business portfolio where your WhatsApp account lives is listed under your company's name or someone else's. In a far from negligible share of SMBs, the account the Instagram campaigns were run from is under the name of an agency, a former employee, or the nephew who helped out three years ago.

If your company is not listed as the owner, do two things today and in this order: ask the current holder to add you as an administrator, and write down who they are, because that is the person the number's registration will depend on. What has to be negotiated by contract so this does not happen again is in who owns your automations.

Into the same folder go the scanned documents Meta asks for to verify the company. You do not need to understand them: they need to be in a PDF and not in a drawer, because the application is submitted on day one and runs in parallel with everything else. Which document is which and how long each stage takes is in how long it takes to implement a WhatsApp chatbot.

The same meeting, with the folder and without it

Take a building-materials yard with two people handling WhatsApp. The business is made up so the difference is visible; the two findings at the end are the kind that always show up.

Without the folder. They say they get "tons of messages". They get quoted a range. The scope ends up written as "answer frequently asked questions and integrate with the management system". Three weeks in, two things surface: the price of cement changes twice a month and nobody writes it down, and the stock lives in a program installed on the PC at the counter.

With the folder. They arrive with 210 monthly conversations, 62 outside business hours and 24 with no answer. Twelve questions with a single answer each, three of them resolved that week among the partners. The list with three written exceptions: volume discount, delivery by zone, store credit account. Escalation to a named person, from 8 a.m. to 6 p.m. And the table that says, in the stock row, "installed program".

That last row is the one that changes the entire project before signing it. There is no live stock lookup; there is a bot that informs, quotes from the written list, and says it will confirm availability during business hours. It is a smaller, cheaper project that does not collapse three months in. The folder did not fix it: it made it visible while there was still a choice.

What changes in the meeting with the folder, and the three things the folder looks like it has and does not

Without the folder, the first meeting is a demo: they show you a bot running and agree to send a proposal. With the folder, it is a diagnosis, and the vendor can close three things that otherwise come out as a range: the platform tier you need, what stays inside the scope, and how many integrations actually have to be built.

The folder also works as a filter, and in both directions. A serious vendor will ask you for these six items on the first call and will tell you which one you are missing. If nobody asks where the stock data comes from or who owns the Meta account, and they still throw you a delivery date, that date is fantasy.

Now, on the other side of that same meeting there is a misunderstanding that costs money: the folder gets you there with a number instead of a range, but it does not build any of the six things it documents. Three of them look so much like the finished product that it is easy to believe you already have them, and the first one is the one most often mistaken for being ready.

Twelve written answers are not a knowledge base. A document with twelve answers answers twelve questions phrased the way you wrote them. What is needed is a system that finds the right answer when the question comes in jumbled or badly written, that knows the exceptions row beats the general list, and that says "I do not know this, let me put you through to someone" instead of filling in the gap. The fastest proof that this leap is not automatic: ask it about the exception, not about the rule. A document pasted into a prompt will quote the list price to the store credit customer, which is exactly the case where the mistake costs money.

The escalation rule on paper is not routing. Getting the right chat to the right person with the full history, without two people answering the same thing and without it getting lost while that person is with a customer, is already a conversation system.

The folder is a snapshot with a date on it. It works the day you put it together. When prices change every week or stock moves every day, the document and reality drift apart, and at that point the problem stops being about content and becomes about the source: either someone owns keeping the data current, with a name and a frequency, or the data has to come from the system where it already changes on its own. Who pays for that every month is a separate decision, and it is in subscription or one-time payment.

Which of the six to put together first if you are not going to do them all

Nobody gathers all six in one afternoon, and the order is not the order of the list. They are ranked by what it costs to fix them late.

First, the ownership of the Meta account. It is fifteen minutes and it is the only one of the six that, if it is wrong, is not fixed with work: it is fixed by losing the number and starting over with another one. Everything else is recoverable.

Second, the data table. That is where you find out whether there is an integration or not, which is the line that moves the budget the most. A folder without this table turns any quote into a three-to-one range, and that range always closes on the same side later.

Third, the written price with its exceptions. It is where the mistake costs money from day one, and it is what no vendor can invent for you.

The other three —the count, the twelve answers, and the escalation rule— can be completed with the project already underway, and no serious vendor will hold you up over them. If you arrive at the first meeting with only those three above, you are ahead of most of the folders that show up complete but with the Meta account under someone's nephew's name.

StriqTech's 15-minute audit is reading whatever you have gathered so far and telling you which one you are missing, which one you do not need in your case, and whether the whole project should wait six months. Write to us at info@striqtech.com or book from striqtech.com/#servicios.